CLASSIFYING MICROWAVE RADAR IMAGES
USING DECISION BASED DATA FUSION
Markus Törmäa, Marcus Engdahlb
Institute of Photogrammetry and Remote Sensing, Helsinki University of Technology, Otakaari 1, Espoo, Finland Markus.Torma@hut.fi
b
Directorate of Earth Observation Programmes, ESA-ESRIN, Via Galileo Galilei, Casella Postale 64, 00044 Frascati, Italy Marcus.Engdahl@esa.int
a
KEY WORDS: Classification, Land Cover, Fusion, SAR
ABSTRACT:
Four different data fusion methods for classifying SAR images with different spatial resolutions were tested and compared. Test area
was Helsinki region in Southern Finland and test data 14 ERS-1/2 Tandem pairs. The best method was a method where a posteriori
probabilities of lower spatial resolution classification are used as a priori probabilities of higher resolution classification. The increase
of overall accuracy was 7-14 %-units depending on date and 10.4% on average when compared to original Tandem pairs. Median
filtering increased classification accuracy, but not that much when data fusion methods were used. This means that the need of spatial
filtering can be at least partially compensated using data fusion of different spatial resolution images.
1. INTRODUCTION
Many governmental institutions have a continuing requirement
to form and implement laws and policies that involve existing
or future land cover and land use. Remote sensing offers means
to provide information about the environment. For example, the
purpose of CORINE Programme is to gather information
relating to the environment for the European Union. In order to
determine and assess the effects of Community’s environment
policy, it is needed to have a proper understanding concerning
the different features of the environment like the state and
geographical distribution of individual environments and natural
areas, the quality and abundance of water resources, land cover
and soil state, and natural hazards (Heymann et.al., 1994).
CORINE land cover classification is produced using optical
satellite images like Landsat ETM. Unfortunately weather
conditions limit the use of optical data. For example, here in
Finland summer is usually quite cloudy, there are usually only
few days when large area of Finland is cloud-free, and during
winter there is dark also daytime. Finnish IMAGE2000 satellite
image mosaic consisted of 36 Landsat ETM images. The target
year was 2000 but only 12 images were taken year 2000 due to
cloudiness (Härmä et.al., 2004).
Due to the relative insensitivity to weather, SAR images are an
interesting alternative to produce information about land cover
and land use. One commercially important application is
cellular network planning, where topographic (DEM) and
morphographic data are needed. SAR interferometry is used to
produce the topographic information. SAR intensity images and
coherence image produced in interferometric process are used to
provide information about the land cover. Morphographic
classes are defined as different land cover types, according to
how they interact and attenuate electromagnetic radiation. The
most common morphographic categories are urban, suburban,
rural, water, open, and forest (Hyyppä et.al., 1999). So, the aim
here is to acquire required spatial information using only two
SAR images taken in slightly different places.
Traditional and difficult problem when classifying SAR images
is speckle, in other words image contains random noise due to
the measurement process. Due to speckle, the statistics of
different land cover classes are rather similar, making the
statistical classification difficult. Speckle can be decreased by
different kind of filters, but developed methods can be difficult
to use due to many parameters that can be unknown. Simple
alternative is performing average filtering. As the area of
averaging increases class histograms start to resemble normal
distribution that is good from statistical pattern recognition
point of view, but spatial resolution decreases at the same time.
This problem is studied from decision based data fusion point of
view in the article. The main principle is that by merging
several different classification results made using images with
different spatial resolutions, it is possible to acquire better
classification than by using any individual classifications. First,
ERS SAR-images are averaged using different sizes of filter and
then these images are classified. Then different methods are
tested to merge these different classifications to form a better
one.
The aims of this research are twofold. First, several rather
simple decision based data fusion methods are introduced and
their performance compared. Second, due to large number of
SAR images, 14 ERS-1/2 Tandem pairs, the effect of
environmental conditions to classification results can be
assessed.
2. DECISION MAKING AND DATA FUSION
Interpretation of remote sensing data can be divided to two
approaches, modelling and classification. Modelling means that
some geophysical parameter is estimated from remote sensing
data, like soil moisture or forest stem volume. The aim of
classification is to divide measurements to discrete groups or
classes according to their similarities. This requires that for each
measurement we make decision about the most proper or likely
class.
2.1 Decision making
A common means to perform classification is to use statistical
pattern recognition framework. There are several different
approaches to classification but the Bayes rule is commonly
utilized. Bayes rule measures the a posteriori probability P(ωj|x)
which feature vector x belongs to class j as (Devivjer and
Kittler, 1982):
P(ω j | x) =
p ( x | ω j ) Pj
p ( x)
,
(1)
where Pj is the a priori probability of class j, p(x|ωj) is the value
of density function of class j and p(x) is the mixture density
function of x defined as
c
p ( x) = ∑ p ( x | ω j ) Pj ,
j =1
(2)
where c is number of classes.
Feature vector x can be classified to class j if the a posteriori
probability of that class is larger than the a posteriori
probabilities of other classes. In that case the decision rule is
called Bayes rule for minimum error. If it is thought that
uncertain classification should not be done, rejection threshold
λr can be used. In that case decision rule becomes
ω ( x) = ωi if
c
P(ωi | x) = max P(ω j | x) ≥ 1 − λr
j =1
(3)
ω ( x) = ω0 if
c
max P(ω j | x) < 1 − λr
j =1
In other words, classification decision is not made if the largest
a posteriori probability is less than 1 - rejection threshold λr.
Cost function λ(ωi|ωj) can be used to make some classification
decisions more important than the others. The purpose of cost
function is to punish wrong decisions so it can be thought to
weight different decisions.
The density function measures the distance between feature
vector x and class j. Remote sensing data is often normally
distributed, especially optical data, so normally distributed
density function
p ( x | ω j ) = (2π ) − d / 2 Σ j
−
1  − 1 ( x − µ ) T Σ −1 ( x − µ ) 
j
j
j 


2 e 2
feature vector is combined directly from different data
sources.
• Feature based fusion means that features have been extracted
from different data sources using e.g. image segmentation. In
this case the features can be e.g. size, shape and average
intensity level of areas. These features form feature vectors
describing the extracted objects.
• Decision based fusion means that the objects have been
identified from individual data sources and then these
interpretation results are combined using e.g. rules to
reinforce common interpretation.
2.3 Decision based data fusion methods
One way to perform decision based data fusion is to use the
class labels of individual classifications and making some kind
of majority decision. This majority decision can be due to
simple majority voting (Ho et.al., 1994) or consensus builder
(Liu et.al., 2002). Other examples of this approach include
methods for making the decision by ranking the class labels of
individual classifications (Ho et.al., 1994), using special neural
network classifier to classify samples if their statistical and
neural network classifications disagree (Kanellopuolos et.al.
1993), or using the classification of expert system as input to
neural network classifier (Liu et.al., 2002). This approach could
be called hard decision based data fusion.
Soft ecision based data fusion is another alternative. In that case
probabilities or other kind of measures which pixel belongs to
certain class are used. This latter approach is used in this study.
All implemented data fusion methods use the output of Bayes
decision rule, i.e. a posteriori probabilities, as their input.
2.3.1 Maximum a posteriori probability: In this case
classification decision is based on the maximum a posteriori
probability of different classifications cl. Decision rule is
ω ( x) = ωi if
cl
c
k =1
j =1
P(ωi | x) = max max P(ω j | xk ).
(5)
This rule corresponds to using fuzzy union operator to combine
different fuzzy sets (Zimmermann, 2001). The drawback is that
information about the reliability of different classifications is
not used.
2.3.2 Maximum joint a posteriori probability: A joint a
posteriori probability of classification decision can be computed
as (Swain, 1978)
(4)
cl
P(ωi | x ) = ∏ P(ωi | xk ),
k =1
(6)
can be used. Σj is the covariance matrix of class j, µj the mean
vector of class j and d is the dimensionality of feature space.
2.2 Alternatives for data fusion
Data fusion can be performed on different levels (Pohl and van
Genderen, 1998):
• Pixel based fusion means that the measurements or measured
physical parameters have been fused. In other words, the
where P(ωi|x) is the a posteriori probability of class j computed
using feature vector xk from data source k. In this case it is
assumed that the individual classifications are independent from
each other, but this is not always the case in real life. Another
drawback is that information about the reliability of different
classifications is not used. This can be incorporated by using
accuracy coefficient αk so that the equation of joint a posteriori
probability is
Nr
Date
Track/Fr
ame
Base
(m)
Air temperature
C°° (9:00 UTC)
Wind speed (direction)
(m/s, deg.)
Snow depth
(cm)
1
2
3
4
5
6
7
8
9
10
11
17/18 July 1995
21/22 August 1995
9/10 September 1995
25/26 September 1995
14/15 October 1995
30/31 October 1995
8/9 January 1996
12/13 February 1996
2/3 March 1996
18/19 March 1996
6/7 April 1996
408/2385
408/2385
179/2385
408/2385
179/2385
408/2385
408/2385
408/2385
179/2385
408/2385
179/2385
-2
-72
-28
239
-221
-49
-29
85
-76
80
37
19.5 17.1
17.1 21.1
11.5 10.3
13.5 11.7
8.6 6.2
5.2 -1.7
-6.5 -5.0
-14.9 -10.8
-4.2 -5.5
-3.2 -4.4
5.5 7.6
8 (140) 5 (190)
3 (330) 5 (240)
6 (30) 3 (30)
8 (190) 10 (210)
6 (290) 1 (90)
2 (260) 5 (340)
2 (120) 3 (170)
5 (100) 1 (120)
4 (10) 3 (50)
3 (350) 1 (110)
4 (220) 2 (130)
12
13
14
22/23 April 1996
15/16 June 1996
20/21 July 1996
408/2385
179/2385
179/2385
-58
48
188
14.7 10.0
15.6 14.9
16.0 14.6
5 (240) 3 (50)
7 (330) 6 (340)
4 (70) 5 (40)
00
00
00
00
00
00
13 13
16 19
32 32
39 38
38 (wet snow)
32 (wet snow)
00
00
00
Precipitation
between
images (mm)
3.4
0
14.9
1.2
0
0.1
0
1.4
0.1
0
0
0
0
0.1
Table 1. The environmental conditions of acquired ERS-1/2 SAR images. Images have been taken approximately 10:35 UTC
(Pulliainen et.al., 2003).
3. SAR IMAGES AND GROUND TRUTH
cl
P(ωi | x) = ∏ P(ωi | xk )α k ,
k =1
(7)
Accuracy coefficient αk can be related to classification accuracy
(e.g. the estimated classification accuracy of x) or spatial
resolution giving lower coefficient to classification that is based
on lower resolution images.
2.3.3 A priori probabilities from lower quality probabilities
or other data: A priori probability of Bayes rule represents our
knowledge about the truth of the hypothesis before we have
analysed the current data. This knowledge could be acquired
using old classification or derived from statistical data. One way
to estimate a priori probability is to compute the a posteriori
probabilities of lower resolution classification and use these as a
priori probabilities in higher resolution classification (Schneider
et.al., 2003, Törmä et.al., 2004). In the case of remote sensing,
this lower resolution could mean that the spatial, radiometric or
spectral resolution is lower or otherwise less suitable for
classification problem at hand.
2.3.4 Dempster-Shafer theory of evidence: The mathematical
theory of evidence is a method to combine different data
sources to provide a joint inference concerning the correct
classification. The idea is to assign a so-called mass of evidence
to various labeling propositions for a feature vector. Each vector
has its own mass of evidence describing the likelihood of
different classes as well as some indication about the
uncertainty about the labeling. For each labeling proposition,
values of support and plausibility are computed. Support is
considered to be the minimum amount of evidence in favor of a
particular labeling for a pixel whereas plausibility is the
maximum possible evidence in favor of the labeling. The
difference between the measures of plausibility and support is
called the evidential interval, the true likelihood that the label
under consideration is correct for the pixel is assumed to lie
somewhere in that interval. Orthogonal sum is used to combine
evidence from different sources. After the orthogonal sum has
been applied, the user can then compute the support for and the
plausibility of each class for a feature vector. The final decision
is based on the support and plausibility. In the simplest case a
maximum support rule can be used (Richards, 1993).
Study site is the Helsinki region in Southern Finland, covering
some 50 km x 50 km area. ERS-1/2 images were acquired
during summer 1995 - summer 1996 consists of 14 Tandem
image pairs. The temporal separation between image
acquisitions is 24 hours in a Tandem pair. Measurement
parameters of ERS are C-band, frequency 5.3 GHz, polarization
VV, incidence angle 23 degrees and spatial resolution about 30
m (Kramer, 1996). Two 5-look backscattered intensity images
and an interferometric coherence image estimated with a 5x5
pixel Gaussian window were produced from each Tandem pair.
All image data was orthorectified to national coordinate system
using an INSAR DEM. Weather conditions within image
acquisition times were also recorded. Table 1 presents the dates
and environmental conditions of image acquisitions. The
intensity and coherence images were scaled to 8-bit integers.
Chosen land cover and use classes are presented in table 2 with
number of ground truth samples per class for different spatial
resolutions. Training areas for classes were determined using
Vantaa city regional map made by Vantaa city authorities and
National Land Use and Forest Classification made by National
Land Survey. Regional map and LUFC were combined to one
map with 20 m pixel. In order to decrease the effect of
georeferencing errors, 2 pixels were removed from borders of
map areas. Then the coordinates of pixels were written to asciifiles and systematically sampled so that the training data of each
class would contain about 1000 pixels at 20 m spatial
resolution.
Class
1. Water
2. Agricultural and other open area
3. Dense forest, stem volume over
100 m3/ha
4. Sparse forest, stem volume 50100 m3/ha
5. Single story houses
6. Multi-story houses
7. Industrial area
SUM
80m
74
39
61
40m
287
252
265
20m
1119
1000
1030
58
261
1000
48
40
62
382
191
164
239
1659
779
705
1008
6641
Table 2. Chosen land cover and use classes with number of
ground truth samples per class for different spatial resolutions.
4. EXPERIMENTS
Statistical classifications were made using Bayes decision rule
with Maximum Likelihood density function estimation method.
Classes were supposed to be normally distributed. Classification
errors were estimated using resubstitution and holdout methods.
In resubstitution method the same set is used as training and test
set and in holdout method data is randomly divided to training
and test sets (Devivjer and Kittler, 1982). In order to decrease
the effect of random division, all classifications were repeated
three times and the mean values of accuracy measures
computed. Because resubtitution method is positively biased
and holdout negatively biased, final error estimate was mean
error of these two estimates. Error matrix was used to compare
the classification results and reference data. Several accuracy
measures like Overall accuracy, Producer’s accuracies of
individual classes, and User’s accuracies of individual classes
were computed from error matrix (Lillesand and Kiefer, 1994).
the spatial resolution decreases at the same time. When the
overall accuracies of 20 and 40 m spatial resolution images are
compared, it is noticed that the difference in accuracies is 0-6
%-units depending on Tandem pair and 2.9 %-units on average.
It seems that differences are largest during summer or winter in
rather dry conditions. In other words, moisture decreases the
differences of classification accuracies obtained using different
spatial resolutions. When 20 and 80 m, or 40 and 80 m images
are compared, the differences are larger. They are on average
7.5 %-units between 20 and 80 m images and 4.6 %-units
between 40 and 80 m images. In these cases there are no clear
correlation between difference and image acquisition times.
Data fusion methods presented in chapter 2.3 were implemented
and used the output of the statistical classifier, i.e. a posteriori
probabilities, as their input:
• Maximum a posteriori probability (DF1) method was
implemented as presented in chapter 2.3.1.
• Maximum joint a posteriori probability (DF2) method has
four different versions according were and how the reliability
measures were used. In case A, each individual classification
has same influence to final decision. In cases B, C and D, the
individual classifications are weighted using the probability
of correct classification or fixed values. In case B, the density
function value of class is weighted by the estimate of the
probability of correct classification that is same as the a
posteriori probability of that class. In case C, the density
function value of class is weighted by the maximum a
posteriori probability of decision. In case D, fixed values
were used and they were 0.5 for 80 m spatial resolution
images, 0.7 for 40 m and 0.9 for 20 m.
• A priori probabilities from lower quality probabilities or other
data (DF3) method used the a posteriori probabilities of lower
spatial resolution classification as a priori probabilities of
higher spatial resolution classification.
• Dempster-Shafer theory of evidence (DF4) method was
implemented according to Richards (1994). There were two
versions differing how the uncertainty of labeling was
defined. Case A used a posteriori probabilities of individual
classifications, as case B had fixed values 0.5 for 80 m spatial
resolution images, 0.7 for 40 m and 0.9 for 20 m.
5. RESULTS
5.1 Effect of spatial resolution to statistical classification
Figure 1 illustrates the effect of spatial resolution to the
classification accuracy. Horizontal axis corresponds to ERS-1/2
Tandem pair (table 2) and vertical axis the overall accuracy.
Solid red line represents the accuracies acquired using 20 m
spatial resolution images, solid blue line with "x" 40 m images,
solid green line with "o" 80 m images, dashed red line 20 m
median filtered images, dashed blue line with "x" 40 m median
filtered images, and dashed green line with "o" 80 m median
filtered images.
From classification accuracy point of view, averaging of SAR
images increases overall accuracies. This is due that the random
component of class statistics is suppressed. The drawback is that
Figure 1. The effect of spatial resolution to
the classification accuracy
The classification accuracies depend quite heavily on
acquisition time and their environmental and weather
conditions. The worst classification accuracies were acquired
using image 4 in relatively windy (wind direction about same in
different dates) conditions and image 11 in wet snow
conditions. The dates of best classification accuracies varied
more and depended on spatial resolution. The best overall
accuracy, 44.9%, for 20 m images was obtained using image 5,
46.6% for 40 m spatial resolution using image 8, and 52.8% for
80 m spatial resolution using image 6.
The median filtering of images increases classification
accuracy. The increase is 2-6 %-units (average 3.8 %-units) in
the case of 20 m spatial resolution images, 2-7 %-units (average
4.3 %-units) in the case of 40 m images and 0-5 %-units
(average 2.9 %-units) in the case of 80 m images. The best
overall accuracy, 47.9%, for 20 m images was obtained using
image 8, 52.3% for 40 m spatial resolution using image 2 and
55.9% for 80 m spatial resolution using image 3.
5.2 Comparison of data fusion methods
Figure 2 illustrates the classification accuracies of different data
fusion methods as function of time. Horizontal axis corresponds
to ERS-1/2 Tandem pair and vertical axis overall accuracy.
Solid red line represents the accuracies of statistical classifier
and 20 m spatial resolution images, solid green line with "x"
data fusion DF1, solid blue line with "o" DF2C, dashed green
line with "x" DF3 and dashed blue line with "o" DF4A.
Data fusion increased classification accuracies in each case. The
best method was DF3, the increase of overall accuracy is 7-14
%-units depending on date and 10.4% on average. DF2C was
the second best; the increase is 4-10 %-units and 7.2 %-units on
average. The increase of accuracy was usually smaller during
spring or autumn. The results of different versions of DF2 were
close to each other, but DF2C was usually the best. The
differences between versions of DF4 were larger. The worst
dates were same as before. The best dates varied between
different methods. The best overall accuracy, 55.0%, was
obtained using DF3 and image 2.
Figure 3. The classification accuracies of different data fusion
methods as function of season.
5.4 Individual classes
Figure 2. The classification accuracies of different data fusion
methods as function of time
Median filtering increased classification accuracy somewhat.
The increase was larger, varying 2-6 %-units, for DF1 and
DF4A and smallest, varying 0-3 % units, for DF3. The best
overall accuracy, 56.5%, was obtained using DF3 and image 2.
This means that the need of spatial filtering can be at least
partially compensated using data fusion of different spatial
resolution images.
5.3 Effect of season
The seasonal comparison was made by computing seasonally
averaged Tandem pairs. Images 1, 2, 13 and 14 were temporally
averaged to form summer Tandem pair, images 3, 4, 5, and 6
autumn Tandem pair, images 7 and 8 winter Tandem pair and
images 9, 10, 11 and 12 spring Tandem pair. Also, all 14
Tandem pairs were temporally averaged to form one Tandem
pair. Figure 3 illustrates the classification accuracies of different
data fusion methods as function of season. Horizontal axis
corresponds to seasonal Tandem pair (1: whole year, 2: autumn,
3: winter, 4: spring and 5: summer) and vertical axis overall
accuracy. Solid red line represents the accuracies of statistical
classifier and 20 m spatial resolution images, solid green line
with "x" data fusion method DF1, solid blue line with "o"
DF2C, dashed green line with "x" DF3 and dashed blue line
with "o" DF4A.
Temporal averaging increased classification accuracy but
surprisingly little. The overall accuracies were about the same
for best seasonal Tandem pairs and median filtered Tandem
pairs for DF1, DF2C and DF4A. The increase was larger in the
case of DF3. When temporal averaging included all Tandem
pairs, accuracies were much better. The best overall accuracy,
66.1%, was obtained using DF2C. In the case of DF3, the
increase of classification accuracy was smaller and actually the
overall accuracy of summer Tandem pair was higher (60.4%)
than the Tandem pair of whole year (58.3%).
The classwise accuracies varied a lot. Table 3 presents the best
producer’s and user’s accuracies for each class, used data fusion
method and corresponding image. Water and agricultural and
other open areas were classified rather well, dense forest and
industrial moderately and other classes rather poorly. The best
data fusion method seems to be DF3. There was some
correlation between season and class accuracies. Producer's
accuracies of class Water were smaller during snow-season than
other seasons. In the case of Agricultural and other open areas,
producer's accuracies were larger during dry snow. The user's
accuracies of classes Water and Dense forest were highest
during autumn. The user's accuracies of classes Agricultural and
other open areas and Single story houses were largest during
winter and Multi-story house the best accuracy was achieved
just after snowmelt.
Class
1. Water
2. Agricultural and
other open area
3. Dense forest,
stem volume over
100 m3/ha
4. Sparse forest,
stem volume 50100 m3/ha
5. Single story
houses
6. Multi-story
houses
7. Industrial area
Producer’s
accuracies
98%, DF3, 5,14
78%, DF3, 10
User’s accuracies
61%, DF3, 8
73%, DF3, 3
44%, DF3, 3
54%, DF3, 8
40%, DF2A, 1
33%, DF3, 9
36%, DF3, 13
56%, DF4A, 12
65%, DF3, 6
57%, DF3, 13
95%, DF1, 6,14
93%, DF1, 9
Table 3. The best producer’s and user’s accuracies for each
class, used data fusion method and corresponding image.
Median filtering increased classwise accuracies somewhat, the
increase varies 1-3 % units depending on accuracy measure and
date. The most notable difference was class Single story house,
in which case user's accuracy increases 12 %-units. The
advantage of medium filtering is that variation of classification
accuracy decreases between worst and best accuracies.
5.5 Mixing of classes
The mixing of classes to other classes was studied using error
matrices. Usually mixing happened following way:
• Water: mixed with Dense forest and Sparse forest
• Agricultural and other open areas: Multi-story houses and
Sparse forest
• Dense forest: Sparse forest and Water
• Sparse forest: Dense forest and Single story houses
• Single story houses: Multi-story houses and Sparse forest
• Multi-story houses: Industrial and Single story houses
• Industrial: Multi-story houses and Single story houses
Ho, T., Hull, J., Srihari, S., 1994. Decision combination in
Multiple Classifier Systems. IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 16, no. 1, January 1994,
pp. 66-75.
As expected, forest classes were mainly mixed with each other
and build-up classes with each other. Main exception was that
single story houses and sparse forest was mixed with each other.
Water was mainly mixed with forest classes. Greatest surprise
was that agricultural and other open areas were quite heavily
mixed with multi-story houses. There were little differences
between different data fusion methods. Largest differences were
between statistical classification of 20 m spatial resolution
images and data fusion methods. In statistical classification,
agricultural and other open areas were more mixed with buildup areas.
Härmä, P., Teiniranta, R., Törmä, M., Repo, R., Järvenpää, E.,
Kallio, M., 2004. The Production of Finnish Corine Land Cover
2000 Classification. International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences, Istanbul,
Turkey, Vol. XXXV, part B4, pp. 1330-1335.
5.6 Merging of classes
The effect of merging of classes from seven classes to four
classes was tested using image 6. Four classes were water,
agricultural and other open areas, forest (classes dense and
sparse forest), and build-up area (classes single story houses,
multi-story houses and industrial). Overall classification
accuracies increased quite a lot, as expected. This increase was
about 20 %-units, from 18.8 %-units in case of statistical
classification of 20 m spatial resolution images to 23.7 %-units
in case of DF2B. The best classification accuracy, 77.7%, was
achieved with DF3. Mixing of classes was mainly as water was
mixed with forest, agricultural with build-up, forest with buildup, and build-up with forest. There were no differences with
different data fusion methods in this sense.
6. CONCLUSIONS
Four different data fusion methods for classifying SAR images
with different spatial resolutions were tested and compared. The
best method was DF3 where a posteriori probabilities of lower
spatial resolution classification are used as a priori probabilities
of higher resolution classification. The increase of overall
accuracy was 7-14 %-units depending on date and 10.4% on
average when compared to original Tandem pairs. The increase
of accuracy was usually smaller during spring or autumn.
Median filtering increased classification accuracy, but not that
much when data fusion methods were used. This means that the
need of spatial filtering can be at least partially compensated
using data fusion of different spatial resolution images.
ACKNOWLEDGMENT
The authors wish to thank ESA for providing the SAR data
through ESA Announcement of Opportunity studies AOTSF.301 and A03-277 as well as the town of Vantaa for the highresolution aerial color orthophotos.
Hyyppä, J., Törmä, M., Engdahl, M., 1999. Deriving
morphographic classification for cellular network planning
using SAR interferometry. IGARSS'99, Hampuri, 28.6.2.7.1999. 1999, IEEE, s. 2176-2178.
Jeon, B., Landgrebe, D., 1999. Decision fusion approach for
multitemporal classification. IEEE Transactions on Geoscience
and Remote Sensing, vol. 37, no. 3, May 1999, pp. 1227-1233.
Kanellopuolos, I., Wilkinson, G., Megier, J., 1993. Integration
of neural network and statistical image classification for land
cover mapping. Proceedings of the International Geoscience
and Remote Sensing Symposium IGARSS’93, Kogakuin
University, Tokyo, Japan, 18-21 August. s. 511-513.
Kramer, H., 1996. Observation of the Earth and Its
Applications. 3rd enlarged edition, Springer.
Lillesand T., Kiefer R., 1994. Remote Sensing and Image
Interpretation. John Wiley & Sons, Inc, 1994.
Liu, X-H., Skidmore, A., Van Oosten, H., 2002. Integration of
classification methods for improvement of land-cover map
accuracy. ISPRS Journal of Photogrammetry and Remote
Sensing, 56, pp. 257-268.
Pohl, C., van Genderen, J., 1998. Multisensor Image Fusion in
Remote Sensing: Concepts, Methods and Applications.
International Journal of Remote Sensing, 19(5), pp. 823-854.
Pulliainen, J., Engdahl, M., Halikainen, M., 2003. Feasibility of
multi-temporal interferometric SAR data for stand-level
estimation of boreal forest stem volume. Remote Sensing of
Environment, vol. 85, pp. 397-409.
Richards, J., 1993. Remote Sensing Digital Image Analysis. 2nd
ed., Springer, 1993.
Schneider, A., Friedl, M., McIver, D., Woodcock, C., 2003.
Mapping urban areas by fusing multiple sources of coarse
resolution remotely sensed data. Photogrammetric Engineering
and Remote Sensing, vol. 69, no. 12, December 2003, pp. 13771386.
Swain, P., 1978. Bayesian classification in a time-varying
environment. IEEE Transactions on Systems, Man, and
Cybernetics, vol. 8, no. 12, pp. 879-883.
REFERENCES
Törmä, M., Patrikainen, N., Luojus, K., Lumme, J., Pyysalo, U.,
2004. Tree Species Classification Using ERS SAR and MODIS
NDVI Images. International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences, Istanbul,
Turkey, Vol. XXXV, part B7, pp. 927-932.
Devivjer P., Kittler J., 1982. Pattern Recognition - A Statistical
Approach. Prentice-Hall, 1982.
Zimmermann, H.-J., 2001. Fuzzy Set Theory and Its
Applications. 4th ed. Kluver Academic, p. 17.